solow-regime 0.2.0

Regime-switching models (Markov switching) for the Solow statistical library
Documentation
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
//! `k`-regime Markov switching autoregression of a given order.
//!
//! ```text
//! y_t = a_{S_t} + sum_{j=1}^p phi_{j,S_t} (y_{t-j} - a_{S_{t-j}}) + e_t,
//!     e_t ~ N(0, sigma^2_{S_t})
//! ```
//!
//! The mean `a`, the autoregressive coefficients `phi`, and optionally the
//! variance switch across regimes. The model conditions on the first `order`
//! observations. Estimation maximises the order-dependent Hamilton-filter
//! log-likelihood.

use ndarray::{Array1, Array2};
use solow_core::error::{Error, Result};

use crate::filter::{hamilton_filter, kim_smoother};
use crate::switching::{
    maximize, ols, steady_state, transform_transition, transition_matrix, transition_param_names,
    untransform_transition, var_pop, MarkovResults,
};

/// A `k`-regime Markov switching autoregression awaiting estimation.
///
/// The mean is a switching constant (`trend = 'c'`). The autoregressive
/// coefficients and the variance may optionally switch.
#[derive(Clone, Debug)]
pub struct MarkovAutoregression {
    /// Full endogenous series (length `nobs + order`).
    endog: Array1<f64>,
    k_regimes: usize,
    order: usize,
    switching_ar: bool,
    switching_variance: bool,
    maxiter: usize,
}

impl MarkovAutoregression {
    /// Build a model. `endog` is the full series; `order >= 1`.
    pub fn new(
        endog: Array1<f64>,
        k_regimes: usize,
        order: usize,
        switching_ar: bool,
        switching_variance: bool,
    ) -> Result<Self> {
        if k_regimes < 2 {
            return Err(Error::Value("k_regimes must be >= 2".into()));
        }
        if order < 1 {
            return Err(Error::Value("order must be >= 1".into()));
        }
        if endog.len() <= order {
            return Err(Error::Value("need more observations than the order".into()));
        }
        Ok(MarkovAutoregression {
            endog,
            k_regimes,
            order,
            switching_ar,
            switching_variance,
            maxiter: 2000,
        })
    }

    fn nobs(&self) -> usize {
        self.endog.len() - self.order
    }

    fn k_trans(&self) -> usize {
        self.k_regimes * (self.k_regimes - 1)
    }

    /// Mean parameters: one switching constant per regime.
    fn k_mean(&self) -> usize {
        self.k_regimes
    }

    fn k_ar(&self) -> usize {
        if self.switching_ar {
            self.k_regimes * self.order
        } else {
            self.order
        }
    }

    fn k_var(&self) -> usize {
        if self.switching_variance {
            self.k_regimes
        } else {
            1
        }
    }

    fn k_params(&self) -> usize {
        self.k_trans() + self.k_mean() + self.k_ar() + self.k_var()
    }

    fn mean_start(&self) -> usize {
        self.k_trans()
    }

    // The reference parameter order is [transition, mean, variance, AR], so the
    // variance block precedes the autoregressive block.
    fn var_start(&self) -> usize {
        self.k_trans() + self.k_mean()
    }

    fn ar_start(&self) -> usize {
        self.k_trans() + self.k_mean() + self.k_var()
    }

    fn param_names(&self) -> Vec<String> {
        let k = self.k_regimes;
        let mut names = transition_param_names(k);
        for i in 0..k {
            names.push(format!("const[{i}]"));
        }
        // Variance precedes autoregressive in the reference layout.
        if self.switching_variance {
            for i in 0..k {
                names.push(format!("sigma2[{i}]"));
            }
        } else {
            names.push("sigma2".to_string());
        }
        if self.switching_ar {
            for i in 0..k {
                for j in 0..self.order {
                    names.push(format!("ar.L{}[{i}]", j + 1));
                }
            }
        } else {
            for j in 0..self.order {
                names.push(format!("ar.L{}", j + 1));
            }
        }
        names
    }

    /// AR coefficient(s) for regime `i` from the constrained parameter vector.
    fn ar_coeffs<'a>(&self, c: &'a Array1<f64>, i: usize) -> &'a [f64] {
        let start = if self.switching_ar {
            self.ar_start() + i * self.order
        } else {
            self.ar_start()
        };
        // SAFETY: owned contiguous constrained parameter vector, so `as_slice()`
        // is always `Some`; the empty-slice fallback would only be reached for a
        // non-contiguous input, which this internal caller never produces.
        &c.as_slice().unwrap_or(&[])[start..start + self.order]
    }

    fn transform(&self, u: &Array1<f64>) -> Array1<f64> {
        let k = self.k_regimes;
        let mut c = u.clone();
        transform_transition(&mut c, u, k);
        // Mean: identity.
        // AR: Monahan stationarity transform per regime block.
        let n_ar_blocks = if self.switching_ar { k } else { 1 };
        for b in 0..n_ar_blocks {
            let start = self.ar_start() + b * self.order;
            let uu: Vec<f64> = (0..self.order).map(|j| u[start + j]).collect();
            let cc = constrain_stationary(&uu);
            for (j, &v) in cc.iter().enumerate() {
                c[start + j] = v;
            }
        }
        // Variance: square.
        for idx in self.var_start()..(self.var_start() + self.k_var()) {
            c[idx] = u[idx] * u[idx];
        }
        c
    }

    fn untransform(&self, c: &Array1<f64>) -> Array1<f64> {
        let k = self.k_regimes;
        let mut u = c.clone();
        untransform_transition(&mut u, c, k);
        let n_ar_blocks = if self.switching_ar { k } else { 1 };
        for b in 0..n_ar_blocks {
            let start = self.ar_start() + b * self.order;
            let cc: Vec<f64> = (0..self.order).map(|j| c[start + j]).collect();
            let uu = unconstrain_stationary(&cc);
            for (j, &v) in uu.iter().enumerate() {
                u[start + j] = v;
            }
        }
        for idx in self.var_start()..(self.var_start() + self.k_var()) {
            u[idx] = c[idx].sqrt();
        }
        u
    }

    /// Conditional log-likelihoods `log f(y_t | S_t, S_{t-1}, ..., S_{t-order})`,
    /// shaped `(k^(order+1), nobs)`, indexed by the joint state with `S_t` most
    /// significant.
    fn conditional_loglik(&self, c: &Array1<f64>) -> Array2<f64> {
        let k = self.k_regimes;
        let order = self.order;
        let nobs = self.nobs();
        let ns = k.pow((order + 1) as u32);
        let mean_start = self.mean_start();
        let var_start = self.var_start();
        let two_pi = std::f64::consts::TAU;
        let mut cll = Array2::<f64>::zeros((ns, nobs));

        for state in 0..ns {
            // Decode joint-state digits: d[0] = S_t, d[m] = S_{t-m}.
            let mut digits = vec![0usize; order + 1];
            let mut idx = state;
            for pos in (0..=order).rev() {
                digits[pos] = idx % k;
                idx /= k;
            }
            let s_t = digits[0];
            let mean_t = c[mean_start + s_t];
            let phi = self.ar_coeffs(c, s_t);
            let variance = if self.switching_variance {
                c[var_start + s_t]
            } else {
                c[var_start]
            };
            for t in 0..nobs {
                // In-sample observation index in the full series.
                let full_t = t + order;
                // Predicted mean conditional on the regime path:
                //   y_t = a_{S_t} + sum_j phi_j (y_{t-j} - a_{S_{t-j}}).
                let mut pred = mean_t;
                for j in 1..=order {
                    let s_lag = digits[j];
                    let mean_lag = c[mean_start + s_lag];
                    pred += phi[j - 1] * (self.endog[full_t - j] - mean_lag);
                }
                let resid = self.endog[full_t] - pred;
                cll[[state, t]] = -0.5 * resid * resid / variance - 0.5 * (two_pi * variance).ln();
            }
        }
        cll
    }

    fn neg_loglike(&self, u: &Array1<f64>) -> Result<f64> {
        let c = self.transform(u);
        let c_s = c
            .as_slice()
            .ok_or_else(|| Error::Value("params must be contiguous".into()))?;
        let p = transition_matrix(c_s, self.k_regimes);
        let init = steady_state(&p)?;
        let cll = self.conditional_loglik(&c);
        let out = hamilton_filter(&init, &p, &cll, self.order);
        Ok(-out.llf)
    }

    /// Starting parameters: OLS of `y_t` on `[1, y_{t-1}, ..., y_{t-p}]`
    /// interpolated across regimes, mirroring the reference.
    fn start_params(&self) -> Result<Array1<f64>> {
        let k = self.k_regimes;
        let order = self.order;
        let nobs = self.nobs();
        let mut c = Array1::<f64>::zeros(self.k_params());
        for j in 0..k {
            for i in 0..(k - 1) {
                c[j * (k - 1) + i] = 1.0 / k as f64;
            }
        }
        // Build design [1, y_{t-1}, ..., y_{t-p}].
        let mut x = Array2::<f64>::zeros((nobs, 1 + order));
        let mut yv = Array1::<f64>::zeros(nobs);
        for t in 0..nobs {
            let full_t = t + order;
            yv[t] = self.endog[full_t];
            x[[t, 0]] = 1.0;
            for j in 1..=order {
                x[[t, j]] = self.endog[full_t - j];
            }
        }
        let beta = ols(&x, &yv)?;
        let resid = &yv - &x.dot(&beta);
        let variance = var_pop(&resid);
        // const term in the reference start params uses beta[:k_exog] (the
        // intercept), interpolated; AR uses beta[k_exog:].
        let intercept = beta[0];
        let mean_start = self.mean_start();
        for i in 0..k {
            let frac = i as f64 / k as f64;
            c[mean_start + i] = intercept * frac;
        }
        if self.switching_ar {
            for i in 0..k {
                let frac = i as f64 / k as f64;
                for j in 0..order {
                    c[self.ar_start() + i * order + j] = beta[1 + j] * frac;
                }
            }
        } else {
            for j in 0..order {
                c[self.ar_start() + j] = beta[1 + j];
            }
        }
        let var_start = self.var_start();
        if self.switching_variance {
            for i in 0..k {
                let t = if k == 1 {
                    0.0
                } else {
                    i as f64 / (k as f64 - 1.0)
                };
                c[var_start + i] = variance / 10.0 + t * (variance - variance / 10.0);
            }
        } else {
            c[var_start] = variance;
        }
        Ok(c)
    }

    /// Fit by maximum likelihood.
    pub fn fit(&self) -> Result<MarkovResults> {
        self.fit_from(None)
    }

    /// Fit, optionally starting from a constrained parameter guess.
    pub fn fit_from(&self, start_constrained: Option<Array1<f64>>) -> Result<MarkovResults> {
        let start_c = match start_constrained {
            Some(c) => c,
            None => self.start_params()?,
        };
        let start_u = self.untransform(&start_c);

        let (xopt, _negll, converged) = maximize(
            &start_u,
            |u| self.neg_loglike(u).unwrap_or(1e10),
            self.maxiter,
        );

        let c = self.transform(&xopt);
        let c_s = c
            .as_slice()
            .ok_or_else(|| Error::Value("params must be contiguous".into()))?;
        let p = transition_matrix(c_s, self.k_regimes);
        let init = steady_state(&p)?;
        let cll = self.conditional_loglik(&c);
        let out = hamilton_filter(&init, &p, &cll, self.order);
        let smoothed = kim_smoother(&p, &out, self.k_regimes, self.order);

        Ok(MarkovResults::new(
            c,
            self.param_names(),
            out.llf,
            self.nobs(),
            p,
            init,
            out.filtered_marginal,
            smoothed,
            converged,
        ))
    }
}

/// Monahan / Barndorff-Nielsen-Schou partial-autocorrelation transform mapping
/// unconstrained reals to stationary AR coefficients, matching scipy /
/// `constrain_stationary_univariate`.
///
/// Step 1: map each unconstrained `u_i` to a partial autocorrelation
/// `r_i = u_i / sqrt(1 + u_i^2)`. Step 2: Durbin–Levinson recursion turns the
/// PACFs into AR coefficients. The reference returns the AR polynomial with the
/// sign convention `y_t = phi_1 y_{t-1} + ...`, which for the single-lag case
/// yields `phi = -r = -u / sqrt(1 + u^2)`.
fn constrain_stationary(u: &[f64]) -> Vec<f64> {
    let n = u.len();
    // Partial autocorrelations in (-1, 1).
    let r: Vec<f64> = u.iter().map(|&x| x / (1.0 + x * x).sqrt()).collect();
    // Monahan's recursion: y[k, i] = y[k-1, i] + r[k] * y[k-1, k-i-1].
    let mut y = vec![vec![0.0f64; n]; n];
    for k in 0..n {
        for i in 0..k {
            y[k][i] = y[k - 1][i] + r[k] * y[k - 1][k - i - 1];
        }
        y[k][k] = r[k];
    }
    // The reference returns -y[n-1, :].
    y[n - 1].iter().map(|&v| -v).collect()
}

/// Inverse of [`constrain_stationary`].
fn unconstrain_stationary(c: &[f64]) -> Vec<f64> {
    let n = c.len();
    let mut y = vec![vec![0.0f64; n]; n];
    // y[n-1, :] = -constrained.
    for i in 0..n {
        y[n - 1][i] = -c[i];
    }
    for k in (1..n).rev() {
        for i in 0..k {
            y[k - 1][i] = (y[k][i] - y[k][k] * y[k][k - i - 1]) / (1.0 - y[k][k] * y[k][k]);
        }
    }
    // r = diagonal; x = r / sqrt(1 - r^2).
    (0..n)
        .map(|k| {
            let r = y[k][k];
            r / (1.0 - r * r).sqrt()
        })
        .collect()
}

#[cfg(test)]
mod ar_transform_tests {
    use super::{constrain_stationary, unconstrain_stationary};

    fn close(a: &[f64], b: &[f64], tol: f64) {
        assert_eq!(a.len(), b.len());
        for (x, y) in a.iter().zip(b.iter()) {
            assert!((x - y).abs() <= tol, "got {x}, want {y}");
        }
    }

    #[test]
    fn order1_matches_reference() {
        // constrain([u]) == -u / sqrt(1+u^2).
        for &u in &[-2.0, -0.5, 0.0, 0.5, 2.0] {
            let c = constrain_stationary(&[u]);
            let want = -u / (1.0 + u * u).sqrt();
            assert!((c[0] - want).abs() < 1e-12);
            let back = unconstrain_stationary(&c);
            assert!((back[0] - u).abs() < 1e-9);
        }
    }

    #[test]
    fn order2_matches_reference() {
        // scipy: constrain([0.7, -0.3]) = [-0.40867915, 0.28734789].
        let c = constrain_stationary(&[0.7, -0.3]);
        close(&c, &[-0.40867915, 0.28734789], 1e-7);
        let back = unconstrain_stationary(&c);
        close(&back, &[0.7, -0.3], 1e-9);
    }
}